Automatic Opioid User Detection From Twitter: Transductive Ensemble Built On Different Meta-graph Based Similarities Over Heterogeneous ...
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18)
Automatic Opioid User Detection From Twitter: Transductive Ensemble Built On
Different Meta-graph Based Similarities Over Heterogeneous Information
Network
Yujie Fan, Yiming Zhang, Yanfang Ye ∗ , Xin Li
Department of Computer Science and Electrical Engineering, West Virginia University, WV, USA
{yf0004,ymzhang}@mix.wvu.edu, {yanfang.ye,xin.li}@mail.wvu.edu
Abstract 13,000 people died from heroin overdose, both reflecting sig-
nificant increase from 2002 [NIDA, 2017]. Opioid addiction
Opioid (e.g., heroin and morphine) addiction has is a chronic mental illness that requires long-term treatment
become one of the largest and deadliest epidemics and care [McLellan et al., 2000]. Although Medication As-
in the United States. To combat such deadly epi- sisted Treatment (MAT) using methadone or buprenorphine
demic, in this paper, we propose a novel frame- has been proven to provide best outcomes for opioid addiction
work named HinOPU to automatically detect opi- recovery, stigma (i.e., bias) associated with MAT has limited
oid users from Twitter, which will assist in sharp- its utilization [Saloner and Karthikeyan, 2015]. Therefore,
ening our understanding toward the behavioral pro- there is an urgent need for novel tools and methodologies to
cess of opioid addiction and treatment. In HinOPU, gain new insights into the behavioral processes of opioid ad-
to model the users and the posted tweets as well diction and treatment.
as their rich relationships, we introduce structured In recent years, the role of social media in biomedical
heterogeneous information network (HIN) for rep- knowledge mining has become more and more important
resentation. Afterwards, we use meta-graph based [Hawn, 2009; Alkhateeb et al., 2011]. Due to the increas-
approach to characterize the semantic relatedness ing use of the Internet, never-ending growth of data is gener-
over users; we then formulate different similari- ated from the social media which offers opportunities for the
ties over users based on different meta-graphs on users to freely share opinions and experiences in online com-
HIN. To reduce the cost of acquiring labeled sam- munities. For example, Twitter, as one of the most popular
ples for supervised learning, we propose a trans- social media platforms, has averaged at 328 million monthly
ductive classification method to build the base clas- active users as of the second quarter of 2017 [Neha, 2017].
sifiers based on different similarities formulated by Many Twitter users are willing to share their experiences of
different meta-graphs. Then, to further improve the using opioids (e.g., “It started with anorexia as I was 11, with
detection accuracy, we construct an ensemble to 13 I was on heroin. At every time I leave heroin I fall into
combine different predictions from different base anorexia”), and perceptions toward MAT (e.g., “I stopped
classifiers for opioid user detection. Comprehen- with heroin bc I’m having the methadone, that is the most
sive experiments on real sample collections from effective treatment.”). Thus, the data from social media may
Twitter are conducted to validate the effectiveness contribute information beyond the knowledge of domain pro-
of HinOPU in opioid user detection by compar- fessionals (e.g., psychiatrists and epidemics researchers) and
isons with other alternate methods. could assist in sharpening our understanding toward the be-
havioral process of opioid addiction and treatment.
To combat the opioid addiction epidemic and promote the
1 Introduction practice of MAT, in this paper, we propose a novel frame-
Opioids are a class of drugs including the illegal drug heroin work named HinOPU, a transductive ensemble classification
and powerful pain relievers by legal prescription, such as model built on different meta-graph based similarities over
morphine and oxycodone [NIDA, 2014]. Opioid addiction heterogeneous information network (HIN), to automatically
has become one of the largest and deadliest epidemics in the detect opioid users from Twitter. To model the users and the
U.S. [NIDA, 2017], which has also turned into a serious na- posted tweets as well as their rich semantic relationships, in
tional crisis that affects public health as well as social and HinOPU, we first introduce a HIN [Sun et al., 2011] for rep-
economic welfare. There was a skyrocketing increase of opi- resentation. To capture the complex relationship (e.g. two
oid related death in the past decade: according to Centers for users are relevant if they have posted tweets that discussed
Disease Control and Prevention (CDC), in 2015, more than the same topic, mentioned same people, and also reposted
32,000 Americans died from opioid drugs overdoses and over same tweets from others), we use meta-graphs [Zhao et al.,
2017] to incorporate higher-level semantics to build up relat-
∗
Corresponding author. edness over users. Different similarities over users can then
3357Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18)
Figure 1: System architecture of HinOPU.
be formulated by different meta-graphs. As obtaining the la- supervised learning, transductive classification method is
beled data (either opioid users or not) from Twitter is both proposed to build the base classifiers; then an ensemble
time-consuming and cost-expensive, to reduce the cost of ac- approach is presented to combine different results from the
quiring labeled samples for supervised learning, we propose a base classifiers to make final predictions. (See Section 3.3.)
transductive classification method to build the base classifiers
based on different similarities formulated by different meta- 3 Proposed Method
graphs; to further improve the detection accuracy, we then
In this section, we introduce the detailed approaches of how
present an ensemble approach to combine different predic-
we represent Twitter users, and how we solve the problem of
tions from different base classifiers for opioid user detection.
opioid user detection based on this representation.
2 System Architecture 3.1 HIN Construction
The system architecture of HinOPU is shown in Figure.1, To detect whether a user is an opioid user or not from Twit-
which consists of the following four major components: ter, besides the users’ posted tweets, the following kinds of
relationships are also considered:
• Data collector and preprocessor. We first develop web
crawling tools to collect opioid-related tweets (i.e., the • R1: To describe the relation of a user and his/her posted
tweets include opioid names or their common street names) tweet, we build the user-post-tweet matrix P where each
as well as users’ profiles from Twitter. For privacy protec- element pi,j ∈ {0, 1} denotes if user i posts tweet j.
tion, UserID is used to anonymize each user. For the col- • R2: To denote the relation that a user appreciates a tweet,
lected tweets, the preprocessor will remove all the links, we generate the user-like-tweet matrix L where each ele-
punctuations, stopwords, and conduct lemmatization using ment li,j ∈ {0, 1} means if user i likes tweet j.
Stanford CoreNLP [Manning et al., 2014]; as street names
can be used to imply opioids (e.g., smack is used to refer • R3: Two users can follow each other in Twitter. To rep-
heroin) and may also have different meanings when given resent such kind of user-user relationship, we generate
different contexts, we will then perform word sense dis- the user-follow-user matrix F where each element fi,j ∈
ambiguation (WSD) [Zhong and Ng, 2010] to decide if a {0, 1} denotes whether user i and user j follow each other.
tweet containing the street name is an opioid-related tweet. • R4: If a tweet includes symbol @ followed by a user name,
• Feature extractor and HIN constructor. A bag-of-words it means that the user is mentioned and talked publicly in
feature vector is extracted to represent each user’s posted the tweet. To describe such tweet-user relationship, we
tweet(s); then the relationships among users, tweets and build the tweet-mention-user matrix A where each ele-
topics will be further analyzed, such as, i) user-post-tweet, ment ai,j ∈ {0, 1} indicates if tweet i mentions user j.
ii) user-like-tweet, iii) user-follow-user, iv) tweet-mention- • R5: A tweet can be a repost of another tweet. To represent
user, v) tweet-RT-tweet, and vi) tweet-contain-topic. Based such relationship between two tweets, we build the tweet-
on these extracted features, a structural HIN will be con- RT-tweet matrix X where element xi,j ∈ {0, 1} denotes
structed. (See Section 3.1.) whether tweet i or tweet j is a repost of the other.
• Meta-graph based similarity builder. In this module, dif- • R6: To represent the relation that a tweet contains a specific
ferent meta-graphs are first built from HIN to capture the topic, we generate the tweet-contain-topic matrix C whose
relatedness between users. Then, we integrate similarity element ci,j ∈ {0, 1} indicates whether tweet i contains
of users’ posted tweets and relatedness depicted by each topic j. In our case, we use Latent Dirichlet Allocation
meta-graph to formulate a set of similarity measures over [Blei et al., 2003] for topic extraction from posted tweets.
users. (See Section 3.2.) In order to depict users, tweets, topics and the rich re-
• Transductive ensemble. Based on different similarities lationships among them (i.e., user-user, user-tweet, tweet-
formulated by different meta-graphs in the previous mod- tweet, tweet-topic relations), it is important to model them in
ule, to reduce the cost of acquiring labeled samples for a proper way so that different kinds of relations can be better
3358Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18)
and easier handled. We introduce how to use HIN to repre- S = (VS , ES , AS , RS , ns , nt ), where VS ⊆ V, ES ⊆ E con-
sent the users by using the features described above. We first strained by AS ⊆ A and RS ⊆ R, respectively.
present the concept related to HIN as below.
Based on the HIN schema displayed in Figure.2, we gener-
Definition 1 A heterogeneous information network (HIN) ate seven meaningful meta-graphs to characterize the related-
[Sun et al., 2011] is a graph G = (V, E) with an entity type ness over users (i.e., SID1–SID7 shown in Figure.3). Differ-
mapping φ: V → A and a relation type mapping ψ: E → R, ent meta-graphs measure the relatedness between two users at
where V denotes the entity set and E is the relation set, A different views. For example, SID1 depicts that two users are
denotes the entity type set and R is the relation type set, and related if they have posted tweets that discussed same topic
the number of entity types |A| > 1 or the number of relation and also mentioned same people; while SID7 describes that
types |R| > 1. The network schema [Sun et al., 2011] for two users are connected if there’re tweets they like that dis-
G, denoted as TG = (A, R), is a graph with nodes as entity cussed same topic, mentioned same people, and also reposted
types from A and edges as relation types from R. same tweet from others. Actually, a meta-path is a special
In our case, we have three entity types (i.e., user, tweet case of a meta-graph. But meta-graph is capable to express
and topic) and six relation types as aforementioned. Based more complex relationship in a convenient way. In Figure.3,
on the definitions above, the network schema for HIN in our the meta-paths PID1–PID8 (left) are the special cases of the
application is shown in Figure. 2. constructed meta-graphs SID1–SID7 (right).
Figure 2: Network schema for HIN.
3.2 Meta-graph Based Similarity
The different types of entities and relations motivate us to use
a machine-readable representation to enrich the semantics of
relatedness among users. To handle this, the concept of meta-
path has been proposed.
Figure 3: Meta-paths (left) and meta-graphs (right). (The symbols
Definition 2 A meta-path [Sun et al., 2011] P is a path de-
in this figure are the abbreviations shown in Figure. 2.)
fined on the graph of network schema TG = (A, R), and is
R1 2 R L R
denoted in the form of A1 −−→ A2 −−→ ... −−→ AL+1 , which To measure the relatedness over users using a particular
defines a composite relation R = R1 · R2 · . . . · RL between meta-graph designed above, we use commuting matrix [Sun
types A1 and AL+1 , where · denotes relation composition op- et al., 2011; Zhao et al., 2017] to compute the counting-based
erator, and L is the length of P. similarity matrix for a meta-graph. Take SID1 as an exam-
An example of a meta-path for users based on HIN ple, the commuting matrix of users computed using SID1 is
post contain
schema shown in Figure.2 is: user −−−→ tweet −−−−−→ MS1 = P[(CCT ) ◦ (AAT )]PT , where P, C, A are the ad-
contain−1 post−1 jacency matrices between two corresponding entity types, ◦
topic −−−−−−→ tweet −−−−→ user, which states that two denotes the Hadamard product of two matrices. MS1 (i, j) de-
users can be connected through their posted tweets contain- notes the number of tweet pairs posted by user i and user j
ing same topics. Although meta-path has been shown to be which contain same topic and also mention same people.
useful for relatedness measure between users, it fails to cap- After characterizing the relatedness between users, we uti-
ture a more complex relationship, e.g., two users have posted lize both content- and relation-based information to measure
tweets that discussed the same topic and also mentioned same the similarity over users: we integrate similarity of users’
people. This calls for a better characterization to handle such posted tweets and relatedness depicted by meta-graph to form
complex relationship. Meta-graph is proposed to use a di- a similarity measure matrix over users. The similarity matrix
rected acyclic graph of entity and relation types to capture over users is denoted as M0 , whose element is a combina-
more complex relationship between two HIN entities. The tion of content-based similarity and meta-graph based relat-
concept of meta-graph is given as follow. edness. We define the similarity matrix M0Sk as: M0Sk (i, j) =
Definition 3 A meta-graph [Zhao et al., 2017] S is a di- [1 + log(MSk (i, j) + 1)] × tSim(i, j), where MSk (i, j) is
rected acyclic graph with a single source node ns and a the relatedness between user i and j under meta-graph Sk ,
single target node nt , defined on a HIN G = (V, E) with tSim(i, j) is the cosine similarity between two bag-of-words
schema TG = (A, R). Formally, a meta-graph is defined as feature vectors generated from user i and j’s posted tweets.
3359Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18)
3.3 Transductive Ensemble Classification in HIN that Yh won’t change since it is clamped in step (2). To get
Base Classifier Yq , we partition B into four sub-matrices:
Compared with inductive classification methods [Lu and
Bhh Bhq
Getoor, 2003] which only use objects with known labels for B=
Bqh Bqq
,
training, transductive classification models [Zhu et al., 2003;
Ye et al., 2011] can utilize the relatedness between objects where Bhh denotes the probabilistic transition matrix for la-
to propagate labels and thus reduce the cost of acquiring la- beled users to labeled users, Bhq is the one for labeled users
beled data. In our case, since it is time-consuming and cost- to unlabeled users - and vise versa for Bqh , and Bqq is the one
expensive to obtain the labeled data (either opioid users or for unlabeled users to unlabeled users. Based on the above
not) from Twitter, we propose to use transductive classifica- algorithm, we have the following iterative propagation rule:
tion in HIN to build the base classifier. The concepts of trans- t+1 t
ductive classification in HIN are introduced as follows. t+1 Yh I 0 Yh
Y = = .
Yt+1 Bqh Bqq Ytq
Sm 4 Given a HIN G = (V, E) with m types of entities
Definition q
V = i=1 Vi , where Vi ∈ Ai (i = 1, ..., m). Suppose V 0 is This equation will converge to [Zhu et al., 2003]:
a subset of V which is with class labels of C = {C1 , ..., Cc }, 0
where c is the number of classes. The task of transductive Yh Yh
Y0 = = lim Yt
= .
classification in HIN [Ji et al., 2010] is to predict the labels Y0q t→∞ (I − Bqq )−1 Bqh Yh
for all the unlabeled entities V − V 0 .
Though we have three types of entities (i.e., user, tweet, Then we have Y0q = (I − Bqq )−1 Bqh Yh , where the element
and topic), we are only interested in classifying one certain
0
yi,j in Y0q is the possibility of user i being in class j. At this
entity (i.e., user), denoted as Au .To solve this problem, topol- point, we can get the class label for each unlabeled user i:
0
ogy shrinking sub-network (TSSN) was proposed in [Li et al., Class(i) = argmaxj∈{1,...,c} {yi,j }.
2016] to perform the transductive classification in HIN. We Ensemble
here further extend the definition of TSSN in our application. Ensemble methods are a popular way to overcome instability
Definition 5 Given a HIN G = (V, E), the topology shrink- and increase performance in many machine learning tasks [Ye
ing sub-network (TSSN) [Li et al., 2016] of a certain entity et al., 2009b]. An ensemble of classifiers is a set of classifiers
type Au derived from a meta-graph S is a graph whose nodes whose individual decisions are combined in some way (e.g.,
consist of only entities of type Au and edges connect enti- by weighted or unweighted voting) to classify new samples,
ties that are related by S. Formally, the TSSN is the graph which is shown to be much more accurate than the individual
GAu ,S = (Vu , EAu , RAu ), where EAu = {ei,j |svi vj ∈ classifiers that make them up [Ye et al., 2009a]. Typically, an
S, vi , vj ∈ Vu } and RAu = {ri,j |ri,j is the weight of ei,j }. ensemble can be decomposed into two components [Ye et al.,
2017]: (1) create base classifiers with necessary accuracy and
In our case, RAu is the similarity matrix M0S formulated
diversity; (2) aggregate all the outputs of the base classifiers
by the meta-graph S in HIN described in Section 3.2.
into a numeric value as the final output of the ensemble.
In transductive classification model, there are two assump-
As described in Section 3.2, different meta-graphs formu-
tions of consistency: (1) smoothness constraint: entities with
late the similarities between users at different views (i.e., with
tight relationship tend to have a high possibility being in the
different semantic meanings). To detect opioid users from
same class; and (2) fitting constraint: the classification re-
Twitter, we first use the above proposed transductive classi-
sults should consist with the prelabeled information. Follow-
fication method in HIN (TCHin) to build the base classifiers
ing these two assumptions, Gaussian Fields Harmonic Func-
based on different similarities formulated by different meta-
tion (GFHF) [Zhu et al., 2003] method was proposed for
graphs (i.e., SID1–SID7 shown in Figure.3); we then use
classification in homogeneous information network. In our
weighted voting (i.e., the weight of k-th classifier wk is deter-
application, given the prelabeled information and a TSSN
mined by its learning accuracy) to aggregate the classification
GAu ,S = (Vu , EAu , RAu ), we propose an approach (named
results from different base classifier to make final prediction.
TCHin) to further extend the GFHF in HIN to assign the la-
The proposed transduction ensemble classification algorithm
bels (either opioid users or non-opioid users) to the unlabeled
(named TEHin) is shown in Algorithm 1.
entities (i.e., users), which works as follow.
Suppose we have h labeled users and q unlabeled users.
Let Y be a (h + q) ∗ c matrix (c is the number of classes), 4 Experimental Results And Analysis
whose element yi,j denotes if user i belongs to class j. Let In this section, we fully evaluate the performance of our de-
Yh be the top h rows of Y and Yq the remaining q rows. To veloped system HinOPU for opioid user detection.
initialize Yq , a random class is assigned to each element in
Yq . Based on the TSSN, D is defined as a diagonal matrix 4.1 Data Collection and Annotation
whose (i, i)-element is equal to the sum of the i-th row of the To obtain the data from Twitter, we develop a set of tools
weight matrix RAu of GAu ,S . Then, a probabilistic transition to crawl the tweets including keywords of opioids (e.g.,
matrix B is calculated via B = D−1 RAu . We then have the heroin, morphine) and their commonly used street names
following algorithm: (1) propagate labels Y = BY; (2) clamp (e.g., smack, morpho), as well as users’ profiles in a period
Yh to the labeled users; (3) repeat until Y converges. Note of time. By the date, 4,537,615 tweets from 4,015,276 users
3360Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18)
Algorithm 1 Transductive Ensemble Classification (TEHin) • BoW (f-1): Each user’s collected tweet(s) is represented as
Input: a bag-of-words feature vector.
h labeled users, q unlabeled users • BoW+Relations (f-2): This augments bag-of-words with
meta-graphs {Sk }7k=1 relations of R1–R6 described in Section 3.1 as flat features.
similarities M0Sk formulated by meta-graphs {Sk }7k=1
• Meta-path (f-3): We combine the generated eight meta-
Output:
paths PID1–PID8 (left of Figure.3) to formulate the relat-
Labels for q unlabeled users
edness over users, where Laplacian scores [He et al., 2006]
1: for k=1 to 7 do are used to learn the weights; then content-based informa-
(k)
2: Build TSSN GAu ,S based on Sk and M0Sk ; tion is incorporated using the method proposed in Section
3: Initialize Y; 3.2 to construct a meta-path based similarity matrix.
4: Build the base classifier using TCHin to get Y0(k)
q ; • Meta-graph (f-4): Similar to the above setting, we utilize
5: end for
the seven meta-graphs SID1–SID7 (right of Figure.3) to
6: Combine the results using weighted voting: Class(i) =
P7 form a meta-graph based similarity matrix.
argmaxj∈{1,...,c} k=1 wk Y0(k) q (i, j).
Second, based on the HIN-based features (meta-path and
meta-graph), we evaluate our proposed learning algorithms
(i.e., TCHin and TEHin) with other learning mechanisms, in-
through Nov. 2007 to Nov. 2017 have been collected and cluding two inductive methods (i.e., Naive Bayes (NB) and
preprocessed in HinOPU. As heroin addiction occupies the Support Vector Machine (SVM)) and a transductive method
majority of today’s opioid addiction and morphine is one of (i.e., Learning with Local and Global Consistency (LLGC)
the most popular legal prescriptions [NIDA, 2014], we first algorithm proposed in [Zhou et al., 2003]). For SVM, we use
study the tweets containing these two keywords and their LibSVM and the penalty is empirically set to be 1,000 while
commonly used street names, as well as the related users. other parameters are set by default.
To obtain the ground truth, seven groups of annotators (i.e.,
35 persons) with knowledge from domain professional (i.e., 4.3 Comparisons
psychiatrist) spent 180 days to label whether the users are Based on DB1 and DB2 , we show the comparison results for
opioid users or not and whether the tweets (including opi- all the above methods for opioid user detection. We randomly
oid names and their street names) are opioid-related tweets or select a portion of the labeled data (ranging from 90% to
not, based on the posted tweets, users’ profiles, and all other 50%) from DB1 and DB2 to simulate the experiments. We
relation-based information by cross-validations. The mutual use F1 measure for evaluations and the experimental results
agreement is above 95%, and only the ones with agreements are illustrated in Table 1.
are retained. The annotated datasets include: From the results, we can see that feature engineering (f-
2: BoW+Relations) helps the performance of machine learn-
• DB1 : This labeled dataset is based on the heroin-related
ing, since the rich semantics encoded in different types of
tweets (i.e., tweets include keywords of heroin or its com-
relations can bring more information. However, the use of
mon street names) and their corresponding users. It con-
this information is simply flat features, i.e., concatenation of
sists of 11,560 users (5,660 are labeled as opioid users
different features altogether, which is less informative than
and 5,900 are non-opioid users) related to 98,610 tweets
the HIN-based features (f-3: meta-path and f-4: meta-graph).
(53,695 are posted by opioid users and 44,915 are posted
Furthermore, the meta-graph feature does perform better than
by non-opioid users).
meta-path. The reason behind this is that meta-graphs are
• DB2 : This labeled dataset is based on the morphine-related more expressive to characterize a complex relatedness over
tweets (i.e., tweets include keywords of morphine or its users than meta-paths.
common street names) and their corresponding users. It Based on the HIN-based features (f-3: meta-path and f-
contains 11,870 users (5,900 are opioid users and 5,970 are 4: meta-graph), we can see that transductive methods (i.e.,
not) related to 74,490 tweets (43,890 are posted by opioid LLGC, TCHin and TEHin) outperform inductive methods
users and 30,600 are posted by non-opioid users). (i.e., NB and SVM) in both datasets especially when labeled
data is insufficient. To put this into perspective, for these
4.2 Baseline Methods two inductive methods, when labeled data decreases from
90% to 50%, their performances greatly compromise (i.e., F1
We validate the performance of our proposed method in drops more than 10%); while the performances of transduc-
HinOPU for opioid users detection by comparisons with dif- tive methods decrease more elegantly as labeled samples de-
ferent groups of baseline methods. crease. This is because transductive methods not only use the
First, we evaluate different types of features for opioid user information of labeled data for prediction but also utilize the
detection from Twitter. Our proposed method is general for relatedness among labeled and unlabeled samples to propa-
HINs. Thus, a natural baseline is to see whether the knowl- gate labels. We also observe that (1) our proposed transduc-
edge we add in should be represented as HIN instead of other tive classification algorithm TCHin outperforms LLGC over
features. Here we compare four types of features: two tra- HIN in opioid user detection under the same experimental set-
ditional features (BoW and BoW+Relations) and two HIN- tings; (2) the ensemble method TEHin achieves better detec-
based features (meta-path and meta-graph). tion performance compared with TCHin, which demonstrates
3361Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18)
Method NB SVM LLGC TCHin TEHin
Settings f-1 f-2 f-3 f-4 f-1 f-2 f-3 f-4 f-3 f-4 f-3 f-4 Alg.1
90% 0.743 0.756 0.776 0.798 0.809 0.818 0.839 0.860 0.840 0.859 0.854 0.872 0.886
DB1 70% 0.696 0.705 0.739 0.760 0.761 0.771 0.791 0.814 0.819 0.830 0.832 0.850 0.863
50% 0.638 0.648 0.674 0.697 0.706 0.716 0.735 0.759 0.798 0.815 0.811 0.832 0.841
90% 0.754 0.766 0.785 0.804 0.805 0.811 0.841 0.861 0.843 0.862 0.858 0.875 0.888
DB2 70% 0.708 0.717 0.738 0.755 0.766 0.779 0.800 0.811 0.819 0.831 0.830 0.849 0.870
50% 0.652 0.660 0.681 0.704 0.703 0.711 0.739 0.760 0.789 0.803 0.802 0.821 0.841
Table 1: Comparisons of different learning methods in opioid user detection when labeled data decreases from 90% to 50%.
TEHin can overcome the instability and improve the detection propose to utilize not only the users’ posted tweets but also
performance compared with individual classifier. the relationships among users and tweets (i.e., user-user, user-
tweet, tweet-tweet, and tweet-topic relations) for opioid user
4.4 Stability and Scalability detection from Twitter; we then present HIN for feature rep-
In this set of experiments, we systematically evaluate the resentations.
stability and scalability of the developed system HinOPU. HIN has been intensively studied in recent years. It has
Figure 4 shows the overall receiver operating characteristic been applied to various applications, such as scientific publi-
(ROC) curves of HinOPU based on 10-fold cross validations: cation network analysis [Sun et al., 2011], social network
HinOPU achieves an impressive 0.923 average TP rate at the analysis for Twitter users [Fan et al., 2017], and malware
0.134 average FP rate in DB1 and 0.931 average TP rate at detection [Hou et al., 2017]. Several studies have already
the 0.131 average FP rate in DB2 . We also further evalu- investigated the use of HIN information for relevance com-
ate the running time of HinOPU with different sizes of the putation, however, most of them only used meta-path [Sun
datasets. Figure 5 shows its scalability, which illustrates that et al., 2011] to measure the similarity. Such simple path
the running time is quadratic to the number of samples. When structure fails to capture a more complex relationship be-
dealing with more data, approximation or parallel algorithms tween two entities. To address this problem, meta-graph
can be developed. From the results and analysis above, for [Zhao et al., 2017] were proposed to measure the proxim-
practical use, our approach is feasible for real application in ity between two entities. In this paper, for opioid user de-
automatic opioid user detection. tection, we take into consideration of different meta-graphs
which characterize the relatedness over Twitter users at dif-
ferent views (i.e., with different semantic meanings). As ac-
quiring the labeled information (either opioid users or not)
from Twitter is both time-consuming and cost-expensive, we
propose to use transductive classification [Ji et al., 2010;
Li et al., 2016] to solve this learning problem in HIN. To com-
bine different results generated from different base transduc-
tive classifiers guided by different meta-graphs, an ensemble
approach is further presented to make final predictions.
6 Conclusion
Figure 4: Stability evaluation. Figure 5: Scalability evaluation.
In this paper, we propose a framework called HinOPU to au-
tomatically detect opioid users from Twitter. In HinOPU, we
first construct a HIN to leverage the information of users,
5 Related Work tweets and topics as well as the rich relationships among
In recent years, the role of social media in biomedical knowl- them, which gives the user a higher-level semantic represen-
edge mining, such as interactive healthcare and drug phar- tation. Then, meta-graph based approach is used to charac-
macology, has become increasingly important. For example, terize the semantic relatedness over users, based on which
based on users’ posted tweets, a machine learning-based con- different similarities over users are formulated. To reduce
cept extraction system ADRMine was introduced for adverse the cost of acquiring labeled samples, a transductive classi-
drug reactions (ADRs) analysis [Nikfarjam et al., 2015]; fication method (TCHin) is proposed to build the base classi-
SVM classifiers based on the content of twitter messages fiers based on different similarities guided by different meta-
were built to find drug users and the potential adverse events graphs; then an ensemble approach (TEHin) is presented to
[Bian et al., 2012]. However, most of these studies merely combine different predictions from the base classifiers for
used content-based features (e.g., posted tweets or messages) opioid user detection. The promising experimental results
for their applications. Actually, the relations among users are based on the real data collections from Twitter demonstrate
also important for targeted user detection. In this paper, we that HinOPU outperforms other alternate methods.
3362Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI-18)
Acknowledgments [Neha, 2017] Kuhar Neha. Effectiveness of social media as
marketing strategy. IJRITCC, 5:942–944, 2017.
This work is partially supported by the U.S. NSF Grant (CNS-
1618629) and WV HEPC Grant (HEPC.dsr.18.5). [NIDA, 2014] NIDA. America’s Addiction to Opi-
oids: Heroin and Prescription Drug Abuse, 2014.
https://goo.gl/HYZAnh.
References
[NIDA, 2017] NIDA. Overdose Death Rates, 2017.
[Alkhateeb et al., 2011] Fadi M Alkhateeb, Kevin A Clau- https://goo.gl/aVXGu2.
son, and David A Latif. Pharmacist use of social media.
[Nikfarjam et al., 2015] Azadeh Nikfarjam, Abeed Sarker,
International Journal of Pharmacy Practice, 19(2):140–
142, 2011. Karen O’Connor, Rachel Ginn, and Graciela Gonzalez.
Pharmacovigilance from social media: mining adverse
[Bian et al., 2012] Jiang Bian, Umit Topaloglu, and Fan Yu. drug reaction mentions using sequence labeling with word
Towards large-scale twitter mining for drug-related ad- embedding cluster features. Journal of the American Med-
verse events. In SHB, pages 25–32. ACM, 2012. ical Informatics Association, 22(3):671–681, 2015.
[Blei et al., 2003] David M Blei, Andrew Y Ng, and [Saloner and Karthikeyan, 2015] Brendan Saloner and
Michael I Jordan. Latent dirichlet allocation. JMLR, Shankar Karthikeyan. Changes in substance abuse treat-
3(Jan):993–1022, 2003. ment use among individuals with opioid use disorders in
[Fan et al., 2017] Yujie Fan, Yiming Zhang, Yanfang Ye, the united states, 2004-2013. The Journal of the American
Wanhong Zheng, et al. Social media for opioid addic- Medical Association, 314(14):1515–1517, 2015.
tion epidemiology: Automatic detection of opioid addicts [Sun et al., 2011] Yizhou Sun, Jiawei Han, Xifeng Yan,
from twitter and case studies. In CIKM, pages 1259–1267. Philip S Yu, and Tianyi Wu. Pathsim: Meta path-based
ACM, 2017. top-k similarity search in heterogeneous information net-
works. VLDB, 4(11):992–1003, 2011.
[Hawn, 2009] Carleen Hawn. Take two aspirin and tweet
me in the morning: how twitter, facebook, and other [Ye et al., 2009a] Yanfang Ye, Lifei Chen, Dingding Wang,
social media are reshaping health care. Health Affairs, Tao Li, Qingshan Jiang, and Min Zhao. Sbmds: an inter-
28(2):361–368, 2009. pretable string based malware detection system using svm
ensemble with bagging. JCV, 5(4):283–293, 2009.
[He et al., 2006] Xiaofei He, Deng Cai, and Partha Niyogi.
[Ye et al., 2009b] Yanfang Ye, Tao Li, Qingshan Jiang,
Laplacian score for feature selection. In NIPS, pages 507–
514, 2006. Zhixue Han, and Li Wan. Intelligent file scoring system
for malware detection from the gray list. In KDD, pages
[Hou et al., 2017] Shifu Hou, Yanfang Ye, Yangqiu Song, 1385–1394. ACM, 2009.
and Melih Abdulhayoglu. Hindroid: An intelligent an- [Ye et al., 2011] Yanfang Ye, Tao Li, Shenghuo Zhu, Wei-
droid malware detection system based on structured het- wei Zhuang, Egemen Tas, Umesh Gupta, and Melih Ab-
erogeneous information network. In KDD, pages 1507– dulhayoglu. Combining file content and file relations for
1515. ACM, 2017. cloud based malware detection. In SIGKDD, pages 222–
[Ji et al., 2010] Ming Ji, Yizhou Sun, Marina Danilevsky, Ji- 230, 2011.
awei Han, and Jing Gao. Graph regularized transductive [Ye et al., 2017] Yanfang Ye, Tao Li, Donald Adjeroh, and
classification on heterogeneous information networks. In S Sitharama Iyengar. A survey on malware detection us-
PKDD, pages 570–586. Springer, 2010. ing data mining techniques. ACM Computing Surveys,
[Li et al., 2016] Xiang Li, Ben Kao, Yudian Zheng, and 50(3):41, 2017.
Zhipeng Huang. On transductive classification in hetero- [Zhao et al., 2017] Huan Zhao, Quanming Yao, Jianda Li,
geneous information networks. In CIKM, pages 811–820. Yangqiu Song, and Dik Lun Lee. Meta-graph based rec-
ACM, 2016. ommendation fusion over heterogeneous information net-
[Lu and Getoor, 2003] Qing Lu and Lise Getoor. Link-based works. In KDD, pages 635–644. ACM, 2017.
classification. In ICML, volume 3, pages 496–503, 2003. [Zhong and Ng, 2010] Zhi Zhong and Hwee Tou Ng. It
[Manning et al., 2014] Christopher D Manning, Mihai Sur- makes sense: A wide-coverage word sense disambigua-
tion system for free text. In ACL (Demo), pages 78–83.
deanu, John Bauer, Jenny Rose Finkel, Steven Bethard,
Association for Computational Linguistics, 2010.
and David McClosky. The stanford corenlp natural lan-
guage processing toolkit. In ACL (Demo), pages 55–60, [Zhou et al., 2003] Dengyong Zhou, Olivier Bousquet,
2014. Thomas Navin Lal, Jason Weston, and Bernhard
Schölkopf. Learning with local and global consistency. In
[McLellan et al., 2000] A Thomas McLellan, David C
NIPS, volume 16, pages 321–328, 2003.
Lewis, Charles P O’brien, and Herbert D Kleber. Drug
dependence, a chronic medical illness: implications for [Zhu et al., 2003] Xiaojin Zhu, Zoubin Ghahramani, and
treatment, insurance, and outcomes evaluation. The Jour- John D Lafferty. Semi-supervised learning using gaussian
nal of the American Medical Association, 284(13):1689– fields and harmonic functions. In ICML, pages 912–919,
1695, 2000. 2003.
3363You can also read
